The cognitive layer that transforms the gap between what an organization says it does and what it actually does into actionable intelligence — governed by humans, traceable over time, and reversible at any moment.
FlowFabric is an enterprise workflow orchestration platform built around a single founding hypothesis: the gap between formal procedures and real field practices is not only a problem to fix — it is also the organization's true operational intelligence.
FlowFabric observes real practices, compares them to official procedures, transforms the gaps into improvement hypotheses, and submits every change to human validation before it reaches production. Nothing is automated without approval. Everything is traceable.
AI proposes. Humans decide.
FlowFabric is the orchestration layer of the Fabric cognitive OS:
┌──────────────────────────────────────────────┐
│ FlowFabric — Orchestration & governance │ ← this repo
├──────────────────────────────────────────────┤
│ BundleFabric — Agent/tool packaging │ ← programs
├──────────────────────────────────────────────┤
│ KFabric — Corpus & knowledge │ ← raw material
└──────────────────────────────────────────────┘
↑ fed by Jean (field observer)
- KFabric stores, indexes, and exposes business knowledge.
- BundleFabric packages each agent/tool with its manifest, dependencies, and trust score.
- FlowFabric orchestrates bundles in governed workflows — this is where human validation lives.
- Jean is the lightweight, event-driven field observer that captures the gap between formal and real.
| Principle | What it means |
|---|---|
| AI proposes, humans decide | No workflow change reaches production without explicit human validation |
| Mandatory provenance | Every recommendation displays its source, date, and confidence level |
| Business logic in the workflow | Rules live in the workflow DSL, not in opaque agents |
| Sandbox before production | Every change is tested before promotion |
| Minimization | Only collect what is strictly necessary for the stated purpose |
| Instrument processes, not people | Jean is the employee's assistant, not a surveillance tool |
| Full data sovereignty | Customers own their data, corpus, and workflows — full export always possible |
Pre-MVP — Strategic foundation phase.
The documentation foundation is complete. Implementation has not started.
| Document | Status |
|---|---|
VISION.md |
✅ v0.2.0 — Full strategic vision (25 sections) |
INVARIANTS.md |
✅ v0.1.0 — Non-negotiable constraints (5 domains, 20 invariants) |
ARCHITECTURE.md |
✅ v0.1.0 — Technical architecture and roadmap (8 lots) |
ARCHITECTURE_SKILLS.md |
✅ v0.1.0 — Skills, technology verdicts, risk matrix (8 domains) |
MVP.md |
✅ v0.1.0 — First use case: supplier invoice exception processing |
VISION_EXEC.md |
✅ v0.1.0 — One-page executive summary |
SECURITY.md |
✅ v0.1.0 — Vulnerability disclosure policy |
- VISION.md — Purpose, vision, mission, promise, principles, Jean, legal framework, layers, metrics
- VISION_EXEC.md — One-page version for decision-makers and partners
- INVARIANTS.md — Signed list of architectural, legal, governance, and product constraints
- ARCHITECTURE.md — Fabric ecosystem, DSL grammar, component map, runtime stack
- ARCHITECTURE_SKILLS.md — Technology verdicts, skill domains, risk matrix, build vs buy
- MVP.md — First deployment: invoice exception processing, KPIs, architecture, ingestion policy
- SECURITY.md — Responsible disclosure, scope, contact
The first FlowFabric deployment targets supplier invoice exception processing — a process that is:
- Highly manual with existing human validation
- Structured enough to instrument without intrusive observation
- Measurable within 90 days
90-day targets: −30% processing time · −20% manual rework · 100% recommendations with provenance · 0 production changes without human validation.
A few non-obvious choices explained:
Why not Temporal? Too complex for solo development. The MVP uses pgqueuer (Python + asyncio + PostgreSQL). Temporal is re-evaluated if workflow duration exceeds a few minutes at scale.
Why not consent as a legal basis for Jean? Consent is generally invalid in an employment context under CNIL/EDPB guidelines (subordination relationship). Legal bases are: legitimate interest, contract performance (strictly necessary to the relationship with the data subject), or legal obligation — determined treatment by treatment.
Why is Jean event-driven, not continuous? Continuous capture is legally disproportionate (CNIL sanction Feb 4, 2025, €40k for 3–15 min screen captures). It is also technically unnecessary: process patterns emerge from transitions and structural actions, not raw streams.
Why is business logic in the workflow, not in the agent? A rule in an agent is opaque, hard to audit, impossible to modify without code changes. A rule in the workflow DSL is visible, versionable, contestable, and transferable.
FlowFabric is in early strategic phase. The codebase does not exist yet.
If you want to contribute to the vision, architecture, or legal framework, open an issue or a discussion. All contributions must respect the invariants defined in INVARIANTS.md.
FlowFabric is designed to be deployed in compliance with GDPR, the EU AI Act, and applicable national labor law. The legal sections in VISION.md and INVARIANTS.md are formulated at the product and architecture level. They must be validated with a DPO and local legal counsel before any production deployment involving employee data.
Apache 2.0 — see LICENSE.
FlowFabric — 2026 — github.com/fbailleux2/flowfabric